Made a couple of updates to my saas foundation template
- inngest installed by default - all apps need background processing pretty much. especially with AI these days (everyone wants some kind of LLM based feature, and those are best kicked off to background jobs)
- invitations can be shared via links, so email setup not required. I previously set it up so that you needed an email service setup to send out invitation links. But its painful to always need that, especially when getting started out, and sign ups for those start around 20 usd/month, so an unneccerary expense as well. Sharing an invite link with someone is just as good I think, and way smoother to get started with
@i_mika_el@nikitabier its one of the annoying of building with AI as well
I feel like as engineers we like deterministic outputs, so when you are "tweaking" prompts to get better results, its just not a fun process I find
@levelsio Could just be related to availability/scarcity also
Those things could just be more expensive, limiting them to rich ppl, vs actual real nutritional benefits
@DustinSnider Nice. Thx for sharing. Yeah I know ppl have got this setup running in prod already. Just been figuring out for myself how to get it set up
Why are you guys on self hosted runners vs something like using github actions?
I've been thinking about how to automate code reviews and deploys via agents.
I think that in order to do that, you need to come with categorizations schemes that allow you to decide how to proceed with next step.
Imagine we want to automate both "approval" of the pull request, as well as whether it can actually be "deployed without human review".
Those are separate because even though a pull request might be considered ready, it might contain high risk updates, in which case you want a human reviewer in the loop.
For first part, you want each issues that is raised in the code review to have a category assigned to it e.g Criticial/Major/Minor/Nit. And anything above certain category, e.g Major and above is Blocking, and the PR cannot be fixed till those are resolved.
Because the agents always come up with issues, so you need a way to cut-off once a PR is "good enough", and so the categories can help with that.
This also allows an agent reviewer and an agent coder, so they have an agreed on "spec" for how to get code over the line. If automating the coding part also, to stop an infinite review cycle, you could hand off to a human if the review cycles goes beyond say three loops.
For the "Deployable" part, the review agent can also assign a "Required Human Approval" flag, based on adding categorizations around "Deployment Risk" of the feature, and "Blast Radius". It can score each of these based on a set of provided categories, and we can use its answers in the build pipeline to see whether we can get this pull request auto-merged or not.
It can be done over time also. So we never auto-merge to start with, but give some time to review how the agent is categorizing pull requests, and maybe updating our review prompt as well. But we can image we will reach a time where we are start to feel confident enough in the agents scoring of the PRs, and at that point we can start letting code be deployed without human review.
The main idea here is that our review agent needs to create categorizations against each issue, and on the PR as a whole, so we have something concrete to make decistions against at each round.
Designer I spoke is now responsible for creating the code for their own designs.
They have been setup with the code base on local (docker, all that)
Claude Code desktop to create their designs and components in the app
Claude code to create the Pull Request
Claude Code to resolve feedback on pull request from the devs.
And that's it done
None one line of code written by the designer. All over chat.
I was saying earlier, code generation is going to start moving over to non dev roles - think designers, product managers, business analysts
And top management are pushing for this heavily as they see this as a massive opportunity to reduce their cost base in terms of needing developers
Convex just raised 57M
This great agentic app builder called Hercules use Convex in the backend
Convex provides a full typescript serveless backend that provides a runtime for funtions, as well as a database.
It also provides file storage and other services.
So that's like a full backend setup, over one service.
Its setup of using functions, queries and mutations means it scales well and builds clean apps even as they get bigger. The agents can understand it easier than maybe traditional systems, and so keep the whole setup looking better.
I think these days the focus needs to be on how can we create environments that are easy for agents to build large, reliable and clean code bases in.
Convex might be one of those setups. Worth looking into I would say.
This is cool. An OpenSource SEO platform that agents can use to find keywords to generate articles.
This is basically how the web is working now right. Agents find keywords, and generating articles, and getting you SEO results
People were liking this globe from @robj3d3 canivibecodeit stats page
So I added one to Drift - the free and simple web analytics website I launched yesterday.
I launched my first app yesterday where I really have barely looked at any if the code
Spent some time on schema design. And did leverage a saas template I created. So the ai basically knew what structure to follow.
And given those 2 things it basically created an entire app where I didn't even really look at the code
It struck me after making it that I've reached the point also where I'm launching without checking the code